TextGrid2multiple_markers — 昇腾 NPU 适配

pip install modelscope onnx onnxruntime torch_npu

# ModelScope SDK 下载
modelscope download --model xiaobaijunya/TextGrid2multiple_markers --local_dir ./TextGrid2multiple_markers

import numpy as np
import onnxruntime as ort

# 加载 ONNX 模型
model_path = "TextGrid2multiple_markers/extracted/TextGrid2oto/HubertFA_model/1218_hfa_model/model.onnx"
session = ort.InferenceSession(model_path, providers=['CPUExecutionProvider'])

# 准备输入 (16kHz 音频)
audio = np.random.randn(1, 16000).astype(np.float32) * 0.01

# 推理
outputs = session.run(None, {'waveform': audio})
cvnt_logits, ph_frame_logits, ph_edge_logits, ctc_logits = outputs

print(f"CV/N/T logits: {cvnt_logits.shape}")
print(f"Phone frame logits: {ph_frame_logits.shape}")
print(f"Phone edge logits: {ph_edge_logits.shape}")
print(f"CTC logits: {ctc_logits.shape}")

import torch
import torch_npu

# HuBERT 算子级 NPU 推理示例
class HuBERTEncoderBlock(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.attn = ...  # Multi-head self-attention
        self.ffn = ...   # Feed-forward with GELU
        self.norm1 = torch.nn.LayerNorm(768)
        self.norm2 = torch.nn.LayerNorm(768)

    def forward(self, x):
        x = x + self.attn(self.norm1(x))
        x = x + self.ffn(self.norm2(x))
        return x

model = HuBERTEncoderBlock().npu()  # 迁移到 NPU
x = torch.randn(1, 100, 768).npu()
output = model(x)  # NPU 推理

@misc{TextGrid2multiple_markers,
  author = {xiaobaijunya},
  title = {TextGrid2multiple_markers - Chinese HuBERT Forced Alignment},
  year = {2025},
  publisher = {ModelScope},
  url = {https://www.modelscope.cn/models/xiaobaijunya/TextGrid2multiple_markers}
}